Tingjun Huang

RIGOR (Rig-Informed Geometry for Omnidirectional Reconstruction) builds 3D maps from 360° construction-site walkthroughs using a frozen feed-forward model and four-view virtual camera rigs. The pipeline combines validated local pose repair, multi-view loop-closure verification and Sim(3) pose-graph optimization before point-cloud fusion.

Assembled an ORCA hand and trained and deployed reinforcement learning for in-hand object reorientation. Completed imitation learning and color-based sorting tasks, followed by a team course project on proximal–distal interphalangeal (PIP–DIP) joint coupling.

Fine-tuned DINOv2 for visual place recognition and distilled its representations into a lightweight Mamba-based model, targeting efficient perception on mobile platforms.

Co-developed category-specific grasping for a Kinova Gen3 Lite arm using GGCNN and YOLOv5, combining object detection with grasp prediction for irregularly shaped objects.

Co-developed monocular visual navigation for differential-drive and Ackermann-steering platforms, using Gazebo and Webots for simulation.

Designed and implemented a social-navigation agent using vision-language models for high-level task reasoning and policy planning, supervised by Prof. Lin Shao.